Sakha-Judge

Data and code release accompanying the paper "Sakha-Judge: A Cross-Family Benchmark for LLM-as-a-Judge Reliability on a Category-0 Language" (Egorov, Humonen).

This is a partial, in-progress release — more will be added.

What's here

• corpus/ — Yakut instruction-tuning corpus components:
  • translated_alpaca_52002.jsonl — Alpaca translated to Yakut (52,002 examples)
  • synthetic_morphology_11438.jsonl — synthetic morphology tasks (11,438 examples)
  • yakut_russian_dictionary_raw_22841.json — Yakut–Russian dictionary source (22,841 entries)
• rubrics/judge_rubric_prompt.md — the grammar/orthography/content rubric prompt used for both LLM judges and human annotators
• human_labels/human_labels_21x9x3.xlsx — human (philologist) ratings, 21 items × 9 candidate models × 3 rubrics
• judge_scores/gemini_pilot/ — Gemini-3.1-Pro judge scores over the candidate set
• judge_scores/qwen36_pilot/ — Qwen-3.6-Plus judge scores (pilot subset)
• code/
  • llm_judge_eval.py — judge harness (OpenRouter-based, rubric scoring, isotonic calibration, Pearson/Spearman/MAE)
  • train_qlora.py — fine-tuning code for the candidate 9B model
  • eval_metrics.py — translation/generation metrics
• translation_benchmark/ — 100-sentence SAH↔RU parallel benchmark + chrF++ evaluation code

Coming later

Claude-Sonnet-4.6 / DeepSeek-V4-Pro judge runs, the English cross-replication set, and the full merged 341,821-example corpus with its train/eval split.